Content Based Recommendations
Overview of the data¶
In [1]:
#import all the necessary packages.
from PIL import Image
import requests
from io import BytesIO
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import warnings
from bs4 import BeautifulSoup
from nltk.corpus import stopwords
from nltk.tokenize import word_tokenize
import nltk
import math
import time
import re
import os
import seaborn as sns
from collections import Counter
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from sklearn.metrics import pairwise_distances
from matplotlib import gridspec
from scipy.sparse import hstack
import plotly
import plotly.figure_factory as ff
from plotly.graph_objs import Scatter, Layout
plotly.offline.init_notebook_mode(connected=True)
warnings.filterwarnings("ignore")
In [33]:
# we have give a json file which consists of all information about
# the products
# loading the data using pandas' read_json file.
data = pd.read_json('tops_fashion.json')
In [34]:
print ('Number of data points : ', data.shape[0], \
'Number of features/variables:', data.shape[1])
In [35]:
# each product/item has 19 features in the raw dataset.
data.columns # prints column-names or feature-names.
Out[35]:
Of these 19 features, we will be using only 6 features.
1. asin ( Amazon standard identification number)
2. brand ( brand to which the product belongs to )
3. color ( Color information of apparel, it can contain many colors as a value ex: red and black stripes )
4. product_type_name (type of the apperal, ex: SHIRT/TSHIRT )
5. medium_image_url ( url of the image )
6. title (title of the product.)
7. formatted_price (price of the product)In [36]:
data = data[['asin', 'brand', 'color', 'medium_image_url', 'product_type_name', 'title', 'formatted_price']]
In [37]:
print ('Number of data points : ', data.shape[0], \
'Number of features:', data.shape[1])
data.head() # prints the top rows in the table.
Out[37]:
[5.1] Missing data for various features.¶
Basic stats for the feature: product_type_name¶
In [38]:
# We have total 72 unique type of product_type_names
print(data['product_type_name'].describe())
# 91.62% (167794/183138) of the products are shirts,
In [39]:
# names of different product types
print(data['product_type_name'].unique())
In [40]:
# find the 10 most frequent product_type_names.
product_type_count = Counter(list(data['product_type_name']))
product_type_count.most_common(10)
Out[40]:
Basic stats for the feature: brand¶
In [41]:
# there are 10577 unique brands
print(data['brand'].describe())
# 183138 - 182987 = 151 missing values.
In [42]:
brand_count = Counter(list(data['brand']))
brand_count.most_common(10)
Out[42]:
Basic stats for the feature: color¶
In [43]:
print(data['color'].describe())
# we have 7380 unique colors
# 7.2% of products are black in color
# 64956 of 183138 products have brand information. That's approx 35.4%.
In [44]:
color_count = Counter(list(data['color']))
color_count.most_common(10)
Out[44]:
Basic stats for the feature: formatted_price¶
In [45]:
print(data['formatted_price'].describe())
# Only 28,395 (15.5% of whole data) products with price information
In [46]:
price_count = Counter(list(data['formatted_price']))
price_count.most_common(10)
Out[46]:
Basic stats for the feature: title¶
In [47]:
print(data['title'].describe())
# All of the products have a title.
# Titles are fairly descriptive of what the product is.
# We use titles extensively in this workshop
# as they are short and informative.
In [48]:
data.to_pickle('pickels/180k_apparel_data')
We save data files at every major step in our processing in "pickle" files.
In [49]:
# consider products which have price information
# data['formatted_price'].isnull() => gives the information
#about the dataframe row's which have null values price == None|Null
data = data.loc[~data['formatted_price'].isnull()]
print('Number of data points After eliminating price=NULL :', data.shape[0])
In [50]:
# consider products which have color information
# data['color'].isnull() => gives the information about the dataframe row's which have null values price == None|Null
data =data.loc[~data['color'].isnull()]
print('Number of data points After eliminating color=NULL :', data.shape[0])
We brought down the number of data points from 183K to 28K.¶
In [51]:
data.to_pickle('pickels/28k_apparel_data')
In [52]:
#To download the images.
'''
from PIL import Image
import requests
from io import BytesIO
for index, row in images.iterrows():
url = row['large_image_url']
response = requests.get(url)
img = Image.open(BytesIO(response.content))
img.save('images/28k_images/'+row['asin']+'.jpeg')
'''
Out[52]:
Remove near duplicate items¶
In [1]:
# read data from pickle file from previous stage
data = pd.read_pickle('pickels/28k_apparel_data')
# find number of products that have duplicate titles.
print(sum(data.duplicated('title')))
# we have 2325 products which have same title but different color
These shirts are exactly same except in size (S, M,L,XL)¶
 :B00AQ4GMCK |
 :B00AQ4GMTS |
 :B00AQ4GMLQ |
 :B00AQ4GN3I |
These shirts exactly same except in color¶
 :B00G278GZ6 |
 :B00G278W6O |
 :B00G278Z2A |
 :B00G2786X8 |
In our data there are many duplicate products like the above examples, we need to de-dupe them for better results.¶
Remove duplicates : Part 1¶
In [102]:
# read data from pickle file from previous stage
data = pd.read_pickle('pickels/28k_apparel_data')
In [103]:
data.head()
Out[103]:
In [104]:
# Remove All products with very few words in title
data_sorted = data[data['title'].apply(lambda x: len(x.split())>4)]
print("After removal of products with short description:", data_sorted.shape[0])
In [105]:
# Sort the whole data based on title (alphabetical order of title)
data_sorted.sort_values('title',inplace=True, ascending=False)
data_sorted.head()
Out[105]:
Some examples of dupliacte titles that differ only in the last few words.¶
Titles 1: 16. woman's place is in the house and the senate shirts for Womens XXL White 17. woman's place is in the house and the senate shirts for Womens M Grey Title 2: 25. tokidoki The Queen of Diamonds Women's Shirt X-Large 26. tokidoki The Queen of Diamonds Women's Shirt Small 27. tokidoki The Queen of Diamonds Women's Shirt Large Title 3: 61. psychedelic colorful Howling Galaxy Wolf T-shirt/Colorful Rainbow Animal Print Head Shirt for woman Neon Wolf t-shirt 62. psychedelic colorful Howling Galaxy Wolf T-shirt/Colorful Rainbow Animal Print Head Shirt for woman Neon Wolf t-shirt 63. psychedelic colorful Howling Galaxy Wolf T-shirt/Colorful Rainbow Animal Print Head Shirt for woman Neon Wolf t-shirt 64. psychedelic colorful Howling Galaxy Wolf T-shirt/Colorful Rainbow Animal Print Head Shirt for woman Neon Wolf t-shirt
In [106]:
indices = []
for i,row in data_sorted.iterrows():
indices.append(i)
In [107]:
import itertools
stage1_dedupe_asins = []
i = 0
j = 0
num_data_points = data_sorted.shape[0]
while i < num_data_points and j < num_data_points:
previous_i = i
# store the list of words of ith string in a, ex: a = ['tokidoki', 'The', 'Queen', 'of', 'Diamonds', 'Women's', 'Shirt', 'X-Large']
a = data['title'].loc[indices[i]].split()
# search for the similar products sequentially
j = i+1
while j < num_data_points:
# store the list of words of jth string in b, ex: b = ['tokidoki', 'The', 'Queen', 'of', 'Diamonds', 'Women's', 'Shirt', 'Small']
b = data['title'].loc[indices[j]].split()
# store the maximum length of two strings
length = max(len(a), len(b))
# count is used to store the number of words that are matched in both strings
count = 0
# itertools.zip_longest(a,b): will map the corresponding words in both strings, it will appened None in case of unequal strings
# example: a =['a', 'b', 'c', 'd']
# b = ['a', 'b', 'd']
# itertools.zip_longest(a,b): will give [('a','a'), ('b','b'), ('c','d'), ('d', None)]
for k in itertools.zip_longest(a,b):
if (k[0] == k[1]):
count += 1
# if the number of words in which both strings differ are > 2 , we are considering it as those two apperals are different
# if the number of words in which both strings differ are < 2 , we are considering it as those two apperals are same, hence we are ignoring them
if (length - count) > 2: # number of words in which both sensences differ
# if both strings are differ by more than 2 words we include the 1st string index
stage1_dedupe_asins.append(data_sorted['asin'].loc[indices[i]])
# if the comaprision between is between num_data_points, num_data_points-1 strings and they differ in more than 2 words we include both
if j == num_data_points-1: stage1_dedupe_asins.append(data_sorted['asin'].loc[indices[j]])
# start searching for similar apperals corresponds 2nd string
i = j
break
else:
j += 1
if previous_i == i:
break
In [108]:
data = data.loc[data['asin'].isin(stage1_dedupe_asins)]
We removed the dupliactes which differ only at the end.¶
In [109]:
print('Number of data points : ', data.shape[0])
In [110]:
data.to_pickle('pickels/17k_apperal_data')
Remove duplicates : Part 2¶
In the previous cell, we sorted whole data in alphabetical order of titles.Then, we removed titles which are adjacent and very similar title But there are some products whose titles are not adjacent but very similar. Examples: Titles-1 86261. UltraClub Women's Classic Wrinkle-Free Long Sleeve Oxford Shirt, Pink, XX-Large 115042. UltraClub Ladies Classic Wrinkle-Free Long-Sleeve Oxford Light Blue XXL TItles-2 75004. EVALY Women's Cool University Of UTAH 3/4 Sleeve Raglan Tee 109225. EVALY Women's Unique University Of UTAH 3/4 Sleeve Raglan Tees 120832. EVALY Women's New University Of UTAH 3/4-Sleeve Raglan Tshirt
In [65]:
data = pd.read_pickle('pickels/17k_apperal_data')
In [66]:
# This code snippet takes significant amount of time.
# O(n^2) time.
indices = []
for i,row in data.iterrows():
indices.append(i)
stage2_dedupe_asins = []
while len(indices)!=0:
i = indices.pop()
stage2_dedupe_asins.append(data['asin'].loc[i])
# consider the first apperal's title
a = data['title'].loc[i].split()
# store the list of words of ith string in a, ex: a = ['tokidoki', 'The', 'Queen', 'of', 'Diamonds', 'Women's', 'Shirt', 'X-Large']
for j in indices:
b = data['title'].loc[j].split()
# store the list of words of jth string in b, ex: b = ['tokidoki', 'The', 'Queen', 'of', 'Diamonds', 'Women's', 'Shirt', 'X-Large']
length = max(len(a),len(b))
# count is used to store the number of words that are matched in both strings
count = 0
# itertools.zip_longest(a,b): will map the corresponding words in both strings, it will appened None in case of unequal strings
# example: a =['a', 'b', 'c', 'd']
# b = ['a', 'b', 'd']
# itertools.zip_longest(a,b): will give [('a','a'), ('b','b'), ('c','d'), ('d', None)]
for k in itertools.zip_longest(a,b):
if (k[0]==k[1]):
count += 1
# if the number of words in which both strings differ are < 3 , we are considering it as those two apperals are same, hence we are ignoring them
if (length - count) < 3:
indices.remove(j)
In [71]:
# from whole previous products we will consider only
# the products that are found in previous cell
data = data.loc[data['asin'].isin(stage2_dedupe_asins)]
In [74]:
print('Number of data points after stage two of dedupe: ',data.shape[0])
# from 17k apperals we reduced to 16k apperals
In [75]:
data.to_pickle('pickels/16k_apperal_data')
# Storing these products in a pickle file.
# candidates who wants to download these files instead.
6. Text pre-processing¶
In [2]:
data = pd.read_pickle('pickels/16k_apperal_data')
# NLTK download stop words. [RUN ONLY ONCE]
# In the temrinal, type these commands
# $python3
# $import nltk
# $nltk.download()
In [3]:
# we use the list of stop words that are downloaded from nltk lib.
stop_words = set(stopwords.words('english'))
print ('list of stop words:', stop_words)
def nlp_preprocessing(total_text, index, column):
if type(total_text) is not int:
string = ""
for words in total_text.split():
# remove the special chars in review like '"#$@!%^&*()_+-~?>< etc.
word = ("".join(e for e in words if e.isalnum()))
# Conver all letters to lower-case
word = word.lower()
# stop-word removal
if not word in stop_words:
string += word + " "
data[column][index] = string
In [5]:
start_time = time.clock()
# we take each title and we text-preprocess it.
for index, row in data.iterrows():
nlp_preprocessing(row['title'], index, 'title')
# we print the time it took to preprocess whole titles
print(time.clock() - start_time, "seconds")
In [6]:
data.head()
Out[6]:
In [8]:
data.to_pickle('pickels/16k_apperal_data_preprocessed')
Stemming¶
In [7]:
from nltk.stem.porter import *
stemmer = PorterStemmer()
print(stemmer.stem('arguing'))
print(stemmer.stem('fishing'))
# We tried using stemming on our titles and it didnot work very well.
[8] Text based product similarity¶
In [4]:
data = pd.read_pickle('pickels/16k_apperal_data_preprocessed')
data.head()
Out[4]:
In [9]:
# Utility Functions which we will use through the rest of the workshop.
#Display an image
def display_img(url,ax,fig):
# we get the url of the apparel and download it
response = requests.get(url)
img = Image.open(BytesIO(response.content))
# we will display it in notebook
plt.imshow(img)
#plotting code to understand the algorithm's decision.
def plot_heatmap(keys, values, labels, url, text):
# keys: list of words of recommended title
# values: len(values) == len(keys), values(i) represents the occurence of the word keys(i)
# labels: len(labels) == len(keys), the values of labels depends on the model we are using
# if model == 'bag of words': labels(i) = values(i)
# if model == 'tfidf weighted bag of words':labels(i) = tfidf(keys(i))
# if model == 'idf weighted bag of words':labels(i) = idf(keys(i))
# url : apparel's url
# we will devide the whole figure into two parts
gs = gridspec.GridSpec(2, 2, width_ratios=[4,1], height_ratios=[4,1])
fig = plt.figure(figsize=(25,3))
# 1st, ploting heat map that represents the count of commonly ocurred words in title2
ax = plt.subplot(gs[0])
# it displays a cell in white color if the word is intersection(lis of words of title1 and list of words of title2), in black if not
ax = sns.heatmap(np.array([values]), annot=np.array([labels]))
ax.set_xticklabels(keys) # set that axis labels as the words of title
ax.set_title(text) # apparel title
# 2nd, plotting image of the the apparel
ax = plt.subplot(gs[1])
# we don't want any grid lines for image and no labels on x-axis and y-axis
ax.grid(False)
ax.set_xticks([])
ax.set_yticks([])
# we call dispaly_img based with paramete url
display_img(url, ax, fig)
# displays combine figure ( heat map and image together)
plt.show()
def plot_heatmap_image(doc_id, vec1, vec2, url, text, model):
# doc_id : index of the title1
# vec1 : input apparels's vector, it is of a dict type {word:count}
# vec2 : recommended apparels's vector, it is of a dict type {word:count}
# url : apparels image url
# text: title of recomonded apparel (used to keep title of image)
# model, it can be any of the models,
# 1. bag_of_words
# 2. tfidf
# 3. idf
# we find the common words in both titles, because these only words contribute to the distance between two title vec's
intersection = set(vec1.keys()) & set(vec2.keys())
# we set the values of non intersecting words to zero, this is just to show the difference in heatmap
for i in vec2:
if i not in intersection:
vec2[i]=0
# for labeling heatmap, keys contains list of all words in title2
keys = list(vec2.keys())
# if ith word in intersection(lis of words of title1 and list of words of title2): values(i)=count of that word in title2 else values(i)=0
values = [vec2[x] for x in vec2.keys()]
# labels: len(labels) == len(keys), the values of labels depends on the model we are using
# if model == 'bag of words': labels(i) = values(i)
# if model == 'tfidf weighted bag of words':labels(i) = tfidf(keys(i))
# if model == 'idf weighted bag of words':labels(i) = idf(keys(i))
if model == 'bag_of_words':
labels = values
elif model == 'tfidf':
labels = []
for x in vec2.keys():
# tfidf_title_vectorizer.vocabulary_ it contains all the words in the corpus
# tfidf_title_features[doc_id, index_of_word_in_corpus] will give the tfidf value of word in given document (doc_id)
if x in tfidf_title_vectorizer.vocabulary_:
labels.append(tfidf_title_features[doc_id, tfidf_title_vectorizer.vocabulary_[x]])
else:
labels.append(0)
elif model == 'idf':
labels = []
for x in vec2.keys():
# idf_title_vectorizer.vocabulary_ it contains all the words in the corpus
# idf_title_features[doc_id, index_of_word_in_corpus] will give the idf value of word in given document (doc_id)
if x in idf_title_vectorizer.vocabulary_:
labels.append(idf_title_features[doc_id, idf_title_vectorizer.vocabulary_[x]])
else:
labels.append(0)
plot_heatmap(keys, values, labels, url, text)
# this function gets a list of wrods along with the frequency of each
# word given "text"
def text_to_vector(text):
word = re.compile(r'\w+')
words = word.findall(text)
# words stores list of all words in given string, you can try 'words = text.split()' this will also gives same result
return Counter(words) # Counter counts the occurence of each word in list, it returns dict type object {word1:count}
def get_result(doc_id, content_a, content_b, url, model):
text1 = content_a
text2 = content_b
# vector1 = dict{word11:#count, word12:#count, etc.}
vector1 = text_to_vector(text1)
# vector1 = dict{word21:#count, word22:#count, etc.}
vector2 = text_to_vector(text2)
plot_heatmap_image(doc_id, vector1, vector2, url, text2, model)
[8.2] Bag of Words (BoW) on product titles.¶
In [12]:
def bag_of_words_model(doc_id, num_results):
# doc_id: apparel's id in given corpus
# pairwise_dist will store the distance from given input apparel to all remaining apparels
# the metric we used here is cosine, the coside distance is mesured as K(X, Y) = <X, Y> / (||X||*||Y||)
# http://scikit-learn.org/stable/modules/metrics.html#cosine-similarity
pairwise_dist = pairwise_distances(title_features,title_features[doc_id])
# np.argsort will return indices of the smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0,len(indices)):
# we will pass 1. doc_id, 2. title1, 3. title2, url, model
get_result(indices[i],data['title'].loc[df_indices[0]], data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], 'bag_of_words')
print('ASIN :',data['asin'].loc[df_indices[i]])
print ('Brand:', data['brand'].loc[df_indices[i]])
print ('Title:', data['title'].loc[df_indices[i]])
print ('Euclidean similarity with the query image :', pdists[i])
print('='*60)
#call the bag-of-words model for a product to get similar products.
bag_of_words_model(931, 20) # change the index if you want to.
# In the output heat map each value represents the count value
# of the label word, the color represents the intersection
# with inputs title.
#tried 12566
#tried 931
In [13]:
from sklearn.feature_extraction.text import CountVectorizer
title_vectorizer = CountVectorizer()
title_features = title_vectorizer.fit_transform(data['title'])
title_features.get_shape() # get number of rows and columns in feature matrix.
# title_features.shape = #data_points * #words_in_corpus
# CountVectorizer().fit_transform(corpus) returns
# the a sparase matrix of dimensions #data_points * #words_in_corpus
# What is a sparse vector?
# title_features[doc_id, index_of_word_in_corpus] = number of times the word occured in that doc
Out[13]:
[8.5] TF-IDF based product similarity¶
In [14]:
tfidf_title_vectorizer = TfidfVectorizer(min_df = 0)
tfidf_title_features = tfidf_title_vectorizer.fit_transform(data['title'])
# tfidf_title_features.shape = #data_points * #words_in_corpus
# CountVectorizer().fit_transform(courpus) returns the a sparase matrix of dimensions #data_points * #words_in_corpus
# tfidf_title_features[doc_id, index_of_word_in_corpus] = tfidf values of the word in given doc
In [15]:
def tfidf_model(doc_id, num_results):
# doc_id: apparel's id in given corpus
# pairwise_dist will store the distance from given input apparel to all remaining apparels
# the metric we used here is cosine, the coside distance is mesured as K(X, Y) = <X, Y> / (||X||*||Y||)
# http://scikit-learn.org/stable/modules/metrics.html#cosine-similarity
pairwise_dist = pairwise_distances(tfidf_title_features,tfidf_title_features[doc_id])
# np.argsort will return indices of 9 smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the 9 smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0,len(indices)):
# we will pass 1. doc_id, 2. title1, 3. title2, url, model
get_result(indices[i], data['title'].loc[df_indices[0]], data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], 'tfidf')
print('ASIN :',data['asin'].loc[df_indices[i]])
print('BRAND :',data['brand'].loc[df_indices[i]])
print ('Eucliden distance from the given image :', pdists[i])
print('='*125)
tfidf_model(12545, 20)
# in the output heat map each value represents the tfidf values of the label word, the color represents the intersection with inputs title
[8.5] IDF based product similarity¶
In [16]:
idf_title_vectorizer = CountVectorizer()
idf_title_features = idf_title_vectorizer.fit_transform(data['title'])
# idf_title_features.shape = #data_points * #words_in_corpus
# CountVectorizer().fit_transform(courpus) returns the a sparase matrix of dimensions #data_points * #words_in_corpus
# idf_title_features[doc_id, index_of_word_in_corpus] = number of times the word occured in that doc
In [17]:
def n_containing(word):
# return the number of documents which had the given word
return sum(1 for blob in data['title'] if word in blob.split())
def idf(word):
# idf = log(#number of docs / #number of docs which had the given word)
return math.log(data.shape[0] / (n_containing(word)))
In [18]:
# we need to convert the values into float
idf_title_features = idf_title_features.astype(np.float)
for i in idf_title_vectorizer.vocabulary_.keys():
# for every word in whole corpus we will find its idf value
idf_val = idf(i)
# to calculate idf_title_features we need to replace the count values with the idf values of the word
# idf_title_features[:, idf_title_vectorizer.vocabulary_[i]].nonzero()[0] will return all documents in which the word i present
for j in idf_title_features[:, idf_title_vectorizer.vocabulary_[i]].nonzero()[0]:
# we replace the count values of word i in document j with idf_value of word i
# idf_title_features[doc_id, index_of_word_in_courpus] = idf value of word
idf_title_features[j,idf_title_vectorizer.vocabulary_[i]] = idf_val
In [19]:
def idf_model(doc_id, num_results):
# doc_id: apparel's id in given corpus
# pairwise_dist will store the distance from given input apparel to all remaining apparels
# the metric we used here is cosine, the coside distance is mesured as K(X, Y) = <X, Y> / (||X||*||Y||)
# http://scikit-learn.org/stable/modules/metrics.html#cosine-similarity
pairwise_dist = pairwise_distances(idf_title_features,idf_title_features[doc_id])
# np.argsort will return indices of 9 smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the 9 smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0,len(indices)):
get_result(indices[i],data['title'].loc[df_indices[0]], data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], 'idf')
print('ASIN :',data['asin'].loc[df_indices[i]])
print('Brand :',data['brand'].loc[df_indices[i]])
print ('euclidean distance from the given image :', pdists[i])
print('='*125)
idf_model(12566,20)
# in the output heat map each value represents the idf values of the label word, the color represents the intersection with inputs title
Text Semantics based product similarity¶
In [24]:
#from gensim.models import Word2Vec
#from gensim.models import KeyedVectors
import pickle
# in this project we are using a pretrained model by google
# its 3.3G file, once you load this into your memory
# it occupies ~9Gb, so please do this step only if you have >12G of ram
'''
model = KeyedVectors.load_word2vec_format('GoogleNews-vectors-negative300.bin', binary=True)
'''
#In case of not having RAM >= 12GB, use the code below.
with open('word2vec_model', 'rb') as handle:
model = pickle.load(handle)
In [25]:
# Utility functions
def get_word_vec(sentence, doc_id, m_name):
# sentence : title of the apparel
# doc_id: document id in our corpus
# m_name: model information it will take two values
# if m_name == 'avg', we will append the model[i], w2v representation of word i
# if m_name == 'weighted', we will multiply each w2v[word] with the idf(word)
vec = []
for i in sentence.split():
if i in vocab:
if m_name == 'weighted' and i in idf_title_vectorizer.vocabulary_:
vec.append(idf_title_features[doc_id, idf_title_vectorizer.vocabulary_[i]] * model[i])
elif m_name == 'avg':
vec.append(model[i])
else:
# if the word in our courpus is not there in the google word2vec corpus, we are just ignoring it
vec.append(np.zeros(shape=(300,)))
# we will return a numpy array of shape (#number of words in title * 300 ) 300 = len(w2v_model[word])
# each row represents the word2vec representation of each word (weighted/avg) in given sentance
return np.array(vec)
def get_distance(vec1, vec2):
# vec1 = np.array(#number_of_words_title1 * 300), each row is a vector of length 300 corresponds to each word in give title
# vec2 = np.array(#number_of_words_title2 * 300), each row is a vector of length 300 corresponds to each word in give title
final_dist = []
# for each vector in vec1 we caluclate the distance(euclidean) to all vectors in vec2
for i in vec1:
dist = []
for j in vec2:
# np.linalg.norm(i-j) will result the euclidean distance between vectors i, j
dist.append(np.linalg.norm(i-j))
final_dist.append(np.array(dist))
# final_dist = np.array(#number of words in title1 * #number of words in title2)
# final_dist[i,j] = euclidean distance between vectors i, j
return np.array(final_dist)
def heat_map_w2v(sentence1, sentence2, url, doc_id1, doc_id2, model):
# sentance1 : title1, input apparel
# sentance2 : title2, recommended apparel
# url: apparel image url
# doc_id1: document id of input apparel
# doc_id2: document id of recommended apparel
# model: it can have two values, 1. avg 2. weighted
#s1_vec = np.array(#number_of_words_title1 * 300), each row is a vector(weighted/avg) of length 300 corresponds to each word in give title
s1_vec = get_word_vec(sentence1, doc_id1, model)
#s2_vec = np.array(#number_of_words_title1 * 300), each row is a vector(weighted/avg) of length 300 corresponds to each word in give title
s2_vec = get_word_vec(sentence2, doc_id2, model)
# s1_s2_dist = np.array(#number of words in title1 * #number of words in title2)
# s1_s2_dist[i,j] = euclidean distance between words i, j
s1_s2_dist = get_distance(s1_vec, s2_vec)
# devide whole figure into 2 parts 1st part displays heatmap 2nd part displays image of apparel
gs = gridspec.GridSpec(2, 2, width_ratios=[4,1],height_ratios=[2,1])
fig = plt.figure(figsize=(15,15))
ax = plt.subplot(gs[0])
# ploting the heap map based on the pairwise distances
ax = sns.heatmap(np.round(s1_s2_dist,4), annot=True)
# set the x axis labels as recommended apparels title
ax.set_xticklabels(sentence2.split())
# set the y axis labels as input apparels title
ax.set_yticklabels(sentence1.split())
# set title as recommended apparels title
ax.set_title(sentence2)
ax = plt.subplot(gs[1])
# we remove all grids and axis labels for image
ax.grid(False)
ax.set_xticks([])
ax.set_yticks([])
display_img(url, ax, fig)
plt.show()
In [26]:
# vocab = stores all the words that are there in google w2v model
# vocab = model.wv.vocab.keys() # if you are using Google word2Vec
vocab = model.keys()
# this function will add the vectors of each word and returns the avg vector of given sentance
def build_avg_vec(sentence, num_features, doc_id, m_name):
# sentace: its title of the apparel
# num_features: the lenght of word2vec vector, its values = 300
# m_name: model information it will take two values
# if m_name == 'avg', we will append the model[i], w2v representation of word i
# if m_name == 'weighted', we will multiply each w2v[word] with the idf(word)
featureVec = np.zeros((num_features,), dtype="float32")
# we will intialize a vector of size 300 with all zeros
# we add each word2vec(wordi) to this fetureVec
nwords = 0
for word in sentence.split():
nwords += 1
if word in vocab:
if m_name == 'weighted' and word in idf_title_vectorizer.vocabulary_:
featureVec = np.add(featureVec, idf_title_features[doc_id, idf_title_vectorizer.vocabulary_[word]] * model[word])
elif m_name == 'avg':
featureVec = np.add(featureVec, model[word])
if(nwords>0):
featureVec = np.divide(featureVec, nwords)
# returns the avg vector of given sentance, its of shape (1, 300)
return featureVec
Average Word2Vec product similarity.¶
In [27]:
doc_id = 0
w2v_title = []
# for every title we build a avg vector representation
for i in data['title']:
w2v_title.append(build_avg_vec(i, 300, doc_id,'avg'))
doc_id += 1
# w2v_title = np.array(# number of doc in courpus * 300), each row corresponds to a doc
w2v_title = np.array(w2v_title)
In [28]:
def avg_w2v_model(doc_id, num_results):
# doc_id: apparel's id in given corpus
# dist(x, y) = sqrt(dot(x, x) - 2 * dot(x, y) + dot(y, y))
pairwise_dist = pairwise_distances(w2v_title, w2v_title[doc_id].reshape(1,-1))
# np.argsort will return indices of 9 smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the 9 smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0, len(indices)):
heat_map_w2v(data['title'].loc[df_indices[0]],data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], indices[0], indices[i], 'avg')
print('ASIN :',data['asin'].loc[df_indices[i]])
print('BRAND :',data['brand'].loc[df_indices[i]])
print ('euclidean distance from given input image :', pdists[i])
print('='*125)
avg_w2v_model(12566, 20)
# in the give heat map, each cell contains the euclidean distance between words i, j
IDF weighted Word2Vec for product similarity¶
In [29]:
doc_id = 0
w2v_title_weight = []
# for every title we build a weighted vector representation
for i in data['title']:
w2v_title_weight.append(build_avg_vec(i, 300, doc_id,'weighted'))
doc_id += 1
# w2v_title = np.array(# number of doc in courpus * 300), each row corresponds to a doc
w2v_title_weight = np.array(w2v_title_weight)
In [30]:
def weighted_w2v_model(doc_id, num_results):
# doc_id: apparel's id in given corpus
# pairwise_dist will store the distance from given input apparel to all remaining apparels
# the metric we used here is cosine, the coside distance is mesured as K(X, Y) = <X, Y> / (||X||*||Y||)
# http://scikit-learn.org/stable/modules/metrics.html#cosine-similarity
pairwise_dist = pairwise_distances(w2v_title_weight, w2v_title_weight[doc_id].reshape(1,-1))
# np.argsort will return indices of 9 smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the 9 smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0, len(indices)):
heat_map_w2v(data['title'].loc[df_indices[0]],data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], indices[0], indices[i], 'weighted')
print('ASIN :',data['asin'].loc[df_indices[i]])
print('Brand :',data['brand'].loc[df_indices[i]])
print('euclidean distance from input :', pdists[i])
print('='*125)
weighted_w2v_model(12566, 20)
#931
#12566
# in the give heat map, each cell contains the euclidean distance between words i, j
Weighted similarity using brand and color.¶
In [31]:
# some of the brand values are empty.
# Need to replace Null with string "NULL"
data['brand'].fillna(value="Not given", inplace=True )
# replace spaces with hypen
brands = [x.replace(" ", "-") for x in data['brand'].values]
types = [x.replace(" ", "-") for x in data['product_type_name'].values]
colors = [x.replace(" ", "-") for x in data['color'].values]
brand_vectorizer = CountVectorizer()
brand_features = brand_vectorizer.fit_transform(brands)
type_vectorizer = CountVectorizer()
type_features = type_vectorizer.fit_transform(types)
color_vectorizer = CountVectorizer()
color_features = color_vectorizer.fit_transform(colors)
extra_features = hstack((brand_features, type_features, color_features)).tocsr()
In [32]:
def heat_map_w2v_brand(sentance1, sentance2, url, doc_id1, doc_id2, df_id1, df_id2, model):
# sentance1 : title1, input apparel
# sentance2 : title2, recommended apparel
# url: apparel image url
# doc_id1: document id of input apparel
# doc_id2: document id of recommended apparel
# df_id1: index of document1 in the data frame
# df_id2: index of document2 in the data frame
# model: it can have two values, 1. avg 2. weighted
#s1_vec = np.array(#number_of_words_title1 * 300), each row is a vector(weighted/avg) of length 300 corresponds to each word in give title
s1_vec = get_word_vec(sentance1, doc_id1, model)
#s2_vec = np.array(#number_of_words_title2 * 300), each row is a vector(weighted/avg) of length 300 corresponds to each word in give title
s2_vec = get_word_vec(sentance2, doc_id2, model)
# s1_s2_dist = np.array(#number of words in title1 * #number of words in title2)
# s1_s2_dist[i,j] = euclidean distance between words i, j
s1_s2_dist = get_distance(s1_vec, s2_vec)
data_matrix = [['Asin','Brand', 'Color', 'Product type'],
[data['asin'].loc[df_id1],brands[doc_id1], colors[doc_id1], types[doc_id1]], # input apparel's features
[data['asin'].loc[df_id2],brands[doc_id2], colors[doc_id2], types[doc_id2]]] # recommonded apparel's features
colorscale = [[0, '#1d004d'],[.5, '#f2e5ff'],[1, '#f2e5d1']] # to color the headings of each column
# we create a table with the data_matrix
table = ff.create_table(data_matrix, index=True, colorscale=colorscale)
# plot it with plotly
plotly.offline.iplot(table, filename='simple_table')
# devide whole figure space into 25 * 1:10 grids
gs = gridspec.GridSpec(25, 15)
fig = plt.figure(figsize=(25,5))
# in first 25*10 grids we plot heatmap
ax1 = plt.subplot(gs[:, :-5])
# ploting the heap map based on the pairwise distances
ax1 = sns.heatmap(np.round(s1_s2_dist,6), annot=True)
# set the x axis labels as recommended apparels title
ax1.set_xticklabels(sentance2.split())
# set the y axis labels as input apparels title
ax1.set_yticklabels(sentance1.split())
# set title as recommended apparels title
ax1.set_title(sentance2)
# in last 25 * 10:15 grids we display image
ax2 = plt.subplot(gs[:, 10:16])
# we dont display grid lins and axis labels to images
ax2.grid(False)
ax2.set_xticks([])
ax2.set_yticks([])
# pass the url it display it
display_img(url, ax2, fig)
plt.show()
In [33]:
def idf_w2v_brand(doc_id, w1, w2, num_results):
# doc_id: apparel's id in given corpus
# w1: weight for w2v features
# w2: weight for brand and color features
# pairwise_dist will store the distance from given input apparel to all remaining apparels
# the metric we used here is cosine, the coside distance is mesured as K(X, Y) = <X, Y> / (||X||*||Y||)
# http://scikit-learn.org/stable/modules/metrics.html#cosine-similarity
idf_w2v_dist = pairwise_distances(w2v_title_weight, w2v_title_weight[doc_id].reshape(1,-1))
ex_feat_dist = pairwise_distances(extra_features, extra_features[doc_id])
pairwise_dist = (w1 * idf_w2v_dist + w2 * ex_feat_dist)/float(w1 + w2)
# np.argsort will return indices of 9 smallest distances
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
#pdists will store the 9 smallest distances
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
#data frame indices of the 9 smallest distace's
df_indices = list(data.index[indices])
for i in range(0, len(indices)):
heat_map_w2v_brand(data['title'].loc[df_indices[0]],data['title'].loc[df_indices[i]], data['medium_image_url'].loc[df_indices[i]], indices[0], indices[i],df_indices[0], df_indices[i], 'weighted')
print('ASIN :',data['asin'].loc[df_indices[i]])
print('Brand :',data['brand'].loc[df_indices[i]])
print('euclidean distance from input :', pdists[i])
print('='*125)
idf_w2v_brand(12566, 5, 5, 20)
# in the give heat map, each cell contains the euclidean distance between words i, j
In [34]:
# brand and color weight =50
# title vector weight = 5
idf_w2v_brand(12566, 5, 50, 20)
Keras and Tensorflow to extract features¶
In [35]:
import numpy as np
from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Dropout, Flatten, Dense
from keras import applications
from sklearn.metrics import pairwise_distances
import matplotlib.pyplot as plt
import requests
from PIL import Image
import pandas as pd
import pickle
In [ ]:
# https://gist.github.com/fchollet/f35fbc80e066a49d65f1688a7e99f069
# Code reference: https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html
# This code takes 40 minutes to run on a modern GPU (graphics card)
# like Nvidia 1050.
# GPU (NVidia 1050): 0.175 seconds per image
# This codse takes 160 minutes to run on a high end i7 CPU
# CPU (i7): 0.615 seconds per image.
#Do NOT run this code unless you want to wait a few hours for it to generate output
# each image is converted into 25088 length dense-vector
# dimensions of our images.
img_width, img_height = 224, 224
top_model_weights_path = 'bottleneck_fc_model.h5'
train_data_dir = 'images2/'
nb_train_samples = 16042
epochs = 50
batch_size = 1
def save_bottlebeck_features():
#Function to compute VGG-16 CNN for image feature extraction.
asins = []
datagen = ImageDataGenerator(rescale=1. / 255)
# build the VGG16 network
model = applications.VGG16(include_top=False, weights='imagenet')
generator = datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode=None,
shuffle=False)
for i in generator.filenames:
asins.append(i[2:-5])
bottleneck_features_train = model.predict_generator(generator, nb_train_samples // batch_size)
bottleneck_features_train = bottleneck_features_train.reshape((16042,25088))
np.save(open('16k_data_cnn_features.npy', 'wb'), bottleneck_features_train)
np.save(open('16k_data_cnn_feature_asins.npy', 'wb'), np.array(asins))
save_bottlebeck_features()
Visual features based product similarity.¶
In [ ]:
#load the features and corresponding ASINS info.
bottleneck_features_train = np.load('16k_data_cnn_features.npy')
asins = np.load('16k_data_cnn_feature_asins.npy')
asins = list(asins)
# load the original 16K dataset
data = pd.read_pickle('pickels/16k_apperal_data_preprocessed')
df_asins = list(data['asin'])
from IPython.display import display, Image, SVG, Math, YouTubeVideo
#get similar products using CNN features (VGG-16)
def get_similar_products_cnn(doc_id, num_results):
doc_id = asins.index(df_asins[doc_id])
pairwise_dist = pairwise_distances(bottleneck_features_train, bottleneck_features_train[doc_id].reshape(1,-1))
indices = np.argsort(pairwise_dist.flatten())[0:num_results]
pdists = np.sort(pairwise_dist.flatten())[0:num_results]
for i in range(len(indices)):
rows = data[['medium_image_url','title']].loc[data['asin']==asins[indices[i]]]
for indx, row in rows.iterrows():
display(Image(url=row['medium_image_url'], embed=True))
print('Product Title: ', row['title'])
print('Euclidean Distance from input image:', pdists[i])
print('Amazon Url: www.amzon.com/dp/'+ asins[indices[i]])
get_similar_products_cnn(12566, 20)